Research On Graph Convolutional Networks Based Node Classification Method And Application

Graphs are a kind of data structure which models a set of entities and their relationships.The expressive power of graphs enables capturing the structural relations among data,and thus allows to harvest more insights compared to analyzing data in isolation.Node classification on graphs has been applied to many fields,including social network analysis,fraud detection,disease prediction and so on.Due to better ability to capture entities dependencies,graph neural networks like graph convolutional networks are widely used in many graph data application scenarios to perform tasks like node classification.Challenges still exist in efficient learning of complex graph data.In this paper,taking account of the Autism Spectrum Disorders(ASD)classification application based on graphs,we focus on mainly the following two problems: how to improve node classification of graph convolutional network in graph data with label noise;and how to combine graph construction with node classification and study them together for more efficient graph construction and node classification.For graph data with noisy labels,we propose a confident graph convolutional network,which introduces confident learning on graph convolutional network for cleaning noisy label.This method can be divided into graph convolutional network component and confident learning component,to do different job as execute classification and clean noise,respectively.These two components working togother for acquiring robust classification.On the basis,we propose an improved method,giving all unlabeled data pseudo labels to improve ability of estimate and clean noise of confident learning component and improve accuracy of the method.Then we evaluate the effectiveness of these two methods on an experimental dataset built by adding label noise on graph dataset of Cora.Finally,we experiment on real data and data with injected label noise of ABIDE I dataset by constructing the ASD population graph.For the unity of graph construction and node classification,considering population graph construction and ASD classification of ABIDE I dataset,we propose a unity graph convolutional network,which is different from traditional methods that construct graphs and train models repeatedly by hands for the best graph construction strategy.The method we proposed assigns a edge weight coefficient representing importance to each feature for edge construction of graph and put parameterized graph construction and graph convolutional network together to optimize classification object for search the best graph construction strategy automatically.Then we experiment UGCN on the ABIDE I dataset,learning graph construction strategy based on combined factors such as resting-state brain functional connectivity similarity,functional Magnetic Resonance Imaging data acquisition site and sex and executing ASD classification,which prove the effectiveness of UGCN.